Jev's skill library earns trust through raw API receipts, not polished demos

Jev is a typed-decision API: you supply structured evidence and a set of named candidates, and it returns a probability-weighted selection rather than free text. This repo is a curated collection of skills, scenario templates, and recorded input/output pairs built around that API, organized so a coding agent can install and use them directly.

The five installable skills divide the work cleanly. jev handles general question design and tool routing. jev-triage sorts and labels records in bulk. jev-documents finds and checks source evidence. jev-eval scores outputs against a rubric. jev-act selects the next legal action in browser, desktop, or game contexts. Together they cover 108 documented scenarios, with 14 saved API call pairs showing exact inputs and outputs.

The recorded examples are the most useful part. A stuck-loop recovery call returns inspect_input with a yes-probability of 0.88 for the stuck question. A completion-evidence check returns claim_supported: false at probability 0.02 when a job was queued but never executed. A document span selection picks the correct invoice email at 0.97 and marks the contradicting claim as contradicted at probability 1. These are not invented examples—the repo links to the raw API receipts.

The pitfalls section is unusually candid. It warns against rerunning until the answer looks right, treating three agreeing calls as independent evidence, and conflating probability with confidence or score. A 20-paired-case context experiment found that fuller evidence enabled more decisions but did not raise accuracy. A community report on a 1,500-email batch notes a quality drop at large batch sizes without claiming a universal row limit.

The install flow is agent-first: you paste a prompt into Codex, Claude Code, or OpenCode, and the agent fetches the install document, checks your environment, and asks whether you want real Jev via OpenRouter or TypeSafe, or an explicitly labeled simulation. The simulation path tags results with jev_called: false and null probabilities—it never silently substitutes.

The ecosystem section lists over 50 community projects, from a Postgres semantic extension requiring superuser access to a ModernBERT-based local alternative. Each entry notes its source type and includes caveats: one benchmark reports 49/49 tasks by majority vote across three attempts, but the repo flags it is not a pure interface ablation.

The core design principle throughout is that Jev chooses and scores; the host agent acts. No recipe here executes a browser action, merges a PR, or sends a message autonomously. That separation is stated repeatedly and is reflected in every recorded example.